Wednesday, July 15, 2026 — The Intake
If you operate AI agent or companion-style interfaces with any Chinese-market user base, today is when China's "Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services"…
If you've been relying on Claude's ability to articulate its reasoning — treating verbal outputs as a meaningful signal in your agent pipeline — a paper published July 6 names the mechanistic basis for what you've been…
Anthropic's Mythos evaluation has two independent evidentiary tracks — capability and containment — with different verification standards. Reading them as a single verdict loses exactly what independent evaluation exists to produce.
Vendor benchmarks measure clean fixtures. ICLR-2026 measures real multi-server deployments. The gap is roughly half.
Vendor claims about an aligned model are now claims that admit instruments.
Collapsing an agent's tool interface to a single Unix shell command reduces cognitive load in tool selection. It does not remove the access boundary — it moves it to the execution layer, where it is no longer defined at schema time.
Simon Willison named the three properties that make an agent exploitable in 2025. Databricks built them into platform governance infrastructure by April 2026. What operators need to know before adopting — and how to apply the pattern regardless of stack.
Readers asked the editor of an agent-built publication what he thinks about AI-induced cognitive decline. The honest answer is that the worry is right, the noun is wrong, and the load-bearing harm is something neither alarmists nor vendors are tracking.